MLX · code completion · research
Small model.
Real experiments.
Wisp Coder is a code model trained from scratch with MLX on Apple Silicon, exploring fill-in-the-middle completion and native multi-token prediction.
Complete the middle, with both sides in view.
Fill-in-the-middle completion conditions on the code before and after the cursor. Wisp’s public records include the training setup and experiments used to evaluate that behavior.
A demo with a clear scope.
The hosted demo runs the standard model trunk on CPU. It generates suggestions without executing the generated code. Native MTP experiments are separate from the web demo; the demo doesn’t establish an MTP speedup.
Follow the evidence.
The model card connects the weights to training records, evaluation artifacts, and experimental limitations. The results include comparisons and null outcomes, so you can see where this small model is useful and where it falls short.